dsh-model-probe

When you connect multiple providers in DSH, a model can silently go dead due to rate limits, timeouts, or deprecation, yet the UI offers no visible cue. You only find out when you pick it, send a request, and get an error. dsh-model-probe closes that gap: it enumerates every model across all registered providers, probes each one through DSH's own LLM runtime, and maintains a persistent Alive / Dead ledger. In every model picker you will see a green or red dot next to each row, and live models are automatically sorted to the top.

The probe policy is deliberately conservative. A manual full sweep retries up to three times; a runtime error triggers two independent 8-token verification requests that never replay the user's original prompt, so there is no risk of duplicated assistant output. Dead models are rechecked in the background every 30 minutes, and a single successful real request instantly flips the status back to Alive. All health state is keyed by provider plus model, persisted locally, and survives both browser and DSH restarts without storing any credentials.

If you routinely juggle several providers and a dozen models in DSH, and want to glance at the composer and immediately know which models are responsive right now, dsh-model-probe is a lightweight, non-intrusive health dashboard that quietly removes that uncertainty from every single request.